Infrastructure atlas
NVIDIA/CPU + GPU superchip

GH200 Grace Hopper

A coherent Grace CPU plus Hopper GPU package that exposes CPU memory as an extension of the GPU memory system.

ASICGraceHoppercoherent memoryNVLink
GH200 Grace Hopper hardware

01 / GPU memory

96 GB HBM3

144 GB HBM3e variant

02 / CPU

Up to 72 Arm cores

Grace CPU

03 / CPU–GPU

900 GB/s

NVLink-C2C

Precision & peak compute

PrecisionPublished peakBasis
FP8 Tensor3,958 TFLOPSHopper GPU · with sparsity
FP16 / BF16 Tensor1,979 TFLOPSHopper GPU · with sparsity

Peak FLOPS and TOPS are theoretical hardware figures. Sparse and dense modes, chip and system totals, and form factors are not interchangeable.

NVIDIA Grace Hopper specifications

Technical specification

Package
GPU
Hopper · up to 96 GB HBM3 or 144 GB HBM3e
CPU
Grace Arm Neoverse V2 · up to 72 cores
CPU memory
Up to 480 / 512 GB LPDDR5X
Coherency
NVLink-C2C hardware-coherent memory
System & software
Scale-up
GH200 NVL2 / NVL32 system options
Addressable memory
GPU can address CPU memory through coherent fabric
Software
CUDA, CUDA-X, Magnum IO, HPC SDK
Best fit
Large memory footprints and tightly coupled AI/HPC